TL;DR: Simon Willison updated his WebRTC audio playground to support the new GPT-Realtime-2 model and custom document injection.
Summary: The updated WebRTC Audio Session playground allows developers to interact with OpenAI's new GPT-Realtime-2 voice model, which features GPT-5-class reasoning. The tool now supports pasting large document contexts directly into the voice session. This helps developers test realtime audio interactions augmented with custom textual reference material.
Why it matters: It demonstrates how to leverage WebRTC voice models with retrieval-augmented generation (RAG) concepts directly in raw browser sessions. AI builders can use these design patterns to construct highly contextual, low-latency voice agents.
Source: simonwillison.net